The use of a neocognitron in an automatic target recognition (ATR) system is described. An image is acquired, edge detected, segmented, and centered on a log-spiral grid using subsystems not discussed in the paper. A conformal transformation is used to map the log-spiral grid to a computation plane in which rotations and scalings are transformed to displacements along the vertical and horizontal axes, respectively. Since the neocognitron can recognize shifted objects, the use of log-spiral images by the neocognitron enables the system to recognize scaled, rotated, and translated objects. Two modifications to prior neocognitron implementations are described. A new weight reinforcement method is introduced which solves a significant training problem for the neocognitron. A method of reducing training time is also introduced which specifies the initial layer of weights in the network. All subsequent layers are trained using unsupervised learning. Simulation results using 32*32 and 64*64 intercontinental ballistic missile (ICBM) images are presented.>

An artificial neural network that self-organizes to recognize various images presented as a training set is described. One application of the network uses multiple functionally disjoint stages to provide pattern recognition that is invariant to translations of the object in the image plane. The general form of the network uses three stages that perform the functionally disjoint tasks of preprocessing, invariance, and recognition. The preprocessing stage is a single layer of processing elements that performs dynamic thresholding and intensity scaling. The invariance stage is a multilayered connectionist implementation of a modified Walsh-Hadamard transform used for generating an invariant representation of the image. The recognition stage is a multilayered self-organizing neural network that learns to recognize the representation of the input image generated by the invariance stage. The network can successfully self-organize to recognize objects without regard to the location of the object in the image field and has some resistance to noise and distortions.>

The authors present a pipeline method for detection and tracking of pixel-sized moving targets with unknown trajectories from a time sequence of highly noisy images. The pipeline target detection algorithm uses the temporal continuity of the smooth trajectories of moving targets and successfully detects and simultaneously tracks all the target trajectories by mapping them from the image sequence onto a single target frame. The pipeline method overcomes the constraint of a straight line trajectory that most other algorithms require for similar tasks. The algorithm is a complete parallel distributed processing type process, and therefore is highly time-efficient-ideal for real-time detection and tracking of arbitrary target trajectories in high-noise environments.>

A pattern-recognition system that self-organizes to recognize objects by shape as part of an integrated visual network (IVN) for autonomous flight control is presented. The system uses a multistaged hierarchical neural network that exhibits insensitivity to the location of the object in the visual field. The network's two layers perform the functionally disjoint tasks of invariance (position normalization) and recognition (identification of the shape). The invariance stage is a multilayered neural network implementation of a modified Walsh-Hadamard transform that generates a representation of the object that is invariant with respect to the object's position. The recognition stage is a modified version of the Fukushima neocognitron that identifies the position-normalized representation by shape. The inclusion of the invariance stage allows reduction of the massively replicated processing structures used for translation invariance in the neocognition. This system offers roughly the same translation invariance capabilities as the neocognitron, with a dramatic reduction in the number of elements and the network's interconnection complexity

This paper presents a pattern recognition system that self-organizes to recognize objects by shape as part of an Integrated Visual Network (IVN) for autonomous flight control. The system uses a multistaged hierarchical neural network that exhibits insensitivity to the location of the object in the visual field. The network's three layers perform the functionally disjoint tasks of preprocessing (dynamic thresholding), invariance (position normalization), and recognition (identification of the shape). The Preprocessing stage uses a single layer of elements to dynamically threshold the grey level input image into a binary image. The Invariance stage is a multilayered neural network implementation of a modified Walsh-Hadamard transform that generates a representation of the object that is invariant with respect to the object’s position. The Recognition stage is a modified version of Fukushima's Neocognitron that identifies the position normalized representation by shape. The inclusion of the Preprocessing and Invariance stages allows reduction of the massively replicated processing structures used for translation invariance in the Neocognitron. This system offers roughly the same translation invariance capabilities as the Neocognitron with a dramatic reduction in the number of elements and the network's interconnection complexity.

A fast tracking algorithm is used to determine displacements of points (pixels) on object surfaces between frames, a sequence of image frames representing a time-varying scene. The displacement is represented by a four-dimension vector whose elements are two-dimensional translation, rotation, rotational angle on the x-y plane, and scaling factor (zoom parameter). The algorithm was developed in conjunction with research on sensors for machine vision with biological vision features. The proposed sensor has two configurations: a uniform rectangular grid and a nonuniform logarithmic spiral grid. The rectangular grid is very well suited to translational motion in 2-D, whereas the logarithmic spiral grid has been proven to be appropriate for rotation and scaling. The algorithm developed utilizes both configurations for estimating 3-D motion.>

An investigation of the effects of various error components in a mobile robot system, which uses a navigation line, was previously reported by the author in 1985. An analysis of the effects of these error components on the navigation line-width determination and obstacle-range determination is presented that provided equations that are used to predict the error resulting from changes in various system parameters. In particular, the change in the rotation angle of the robot sensor about its optical axis, called the roll angle, is examined. Numerical examples showing the characteristics of the derived error equations are also given.>

An algorithm for translational motion prediction is presented. The algorithm predicts motion without using correspondence and using only one line correlation instead of full image correlation. The algorithm was developed in conjunction with research on sensors for machine vision with biological vision features. Tests with noisy synthetic images produced good results.

Experimental results are presented that support theory published in the literature concerning the use of a navigation line for the guidance of mobile robots. A method for the determination of a robot's position is developed. A specialized edge operator, which aids in the segmentation of a navigation line from an image of a robot's environment, is given. Use of this specialized edge operator in conjunction with the Hough transform is also presented. These methods are used to verify the analytical results given previously. Comparisons are made between experimental data and expected results as a function of various system parameters. Real-time implementation of these methods is considered.

A continuous, straight-edged line is used for the visual navigation of an autonomous mobile robot in a factor environment. This line, which resides on the floor and contrasts with background, may also be used to determine range information. Two methods are developed for determining the range of an object in the sensor's field of view. The effects of various error conditions in the system geometry on each ranging method are determined. Equations are derived which yield the percent error in calculating ranges given estimates of these error conditions. Numerical examples using typical sensor parameters are given.

The use of a contrasting line for the visual navigation of autonomous mobile robots in a factory environment is developed. Minimum and maximum linewidths are determined analytically by considering sensor geometry, field of view, and error conditions present in the system. The effects of these error conditions on the width of the line, as seen in the image plane, determines the optimal linewidth. Numerical examples using typical sensor parameters are given.

Research on the semiautonomous operation of mobile robots in typical pathways is described. The image of the pathway will consist of two nearly vertical lines bounding a region with little texture (the pathway) after correction for perspective. In order to identify pathway boundaries, regions in the image space are examined using an edge detection algorithm, edges between regions are determined by the algorithm, and those corresponding to straight or nearly straight lines with large slope (path boundaries) are identified by means of the Hough transform. Once the path boundaries are identified, the horizontal distance from camera to road edge is determined. Next, a method to detect cubics in the roadway (i.e., obstacles) is presented. The region of interest in the roadway (from the camera to some predetermined distance in front of it) is known from the path boundary algorithm. The interior of this region is examined for edges. If edges are detected, it means that obstacles or shadows are present. A method to separate obstacles from shadows using stere vision is then presented.

Implementation of a three-dimensional video tracking algorithm using CCD's is presented. The algorithm does not require conventional pattern recognition or continuous segmentation updating. The only requirements, for the class of targets presented, are that the sampling lattice contain only points inside the target and that the background vary slowly. Several well-known segmentation techniques are available for initial segmentation and determination of target points. The use of CCD devices simplifies the architecture of the processor and permits real time processing at a high sampling rate. Some of the vector and matrix operations require either the use of continuously variable tapweight transversal filters or analog multipliers. The use of analog multipliers, which can be on-chip, reduces the number of components and offers higher accuracy than presently available variable tap weight transversal filters.